**4. Data and methodology**

All series examined in this studyindustrial production index, yields of 10-year government bond and 3-month Treasury bill, FTSE-Kuala Lumpur Composite Index, broad money (M2) and exportsare collected from Thomson-Reuters Datastream. The data are monthly data and span the time period from January 1996 to December 2016. The industrial production index is used to signal economic growth, while the yield spread is calculated as the difference between the yields of 10-year government bond and 3-month Treasury bill.

In order to control for other leading variables that could be influencing economic growth, a single factor is constructed to extract the common signals from all the leading indicators. In this study, the leading indicators considered are FTSE-Kuala Lumpur Composite Index, broad money (M2), and exports value. It is also important to point out that the inclusion of all these leading indicators into the regression model could lead to multicollinearity issue, it is best to express them as a single controlled factor. Besides, the main objective of this study is to examine the cointegration of the yield spread and growth, without putting much emphasis on other leading economic variables. Hence, the principal component analysis technique is used to extract the single factor from the movements of these indicators.

In terms of methodology, the relationship between the yield spread and economic growth is established using the autoregressive distributed lag (ARDL) framework by Pesaran and Shin [28], Pesaran et al. [29], and Pesaran [30]. Given the characteristics of the cyclical components of the data, applying the conventional Granger [11] and, Engle and Granger [12] cointegration technique is not applicable in cases of variables that are integrated of different orders. This is because, a prerequisite for applying the abovementioned cointegration analysis is that all timeseries are nonstationary and must be of the same order. The ARDL method stands out among other regression methods as it does not involve pre-testing variables, which means that testing on the existence relationship between variables that are integrated of different order of purely I(0), purely I(1), or mixture of both (Duasa [31], Nkoro and Uko [32]). Another superior feature of the ARDL method is that it avoids the larger number of specifications to be made into the regressions, particularly with regard to the decisions on the number of endogeneous and exogenous variables (if any) to be included, as well as the optimal number of lags to be specified. With ARDL, it is possible that different variables have different optimal lags, which is impossible on the application of the standard cointegration test. Most importantly, the model could be used with limited sample data (30–80 observations) since in our analysis, even though the total observations for the whole sample is 249, the sub-sample of the data will be much smaller in quantity. On top of that, the ARDL method will automatically run models with different lags to choose the best estimation model based on the specified selection criteria, in our case, using Akaike info criterion (AIC).

In order to establish the relationship between the yield spread and growth, a linear regression relating the future growth to the current values of the yield spread is considered. The model below is developed based on the model used by Stock and Watson [17] and Zulkhibri and Abdul Rani [10], which has been extended to

*Has the Yield Curve Accurately Predicted the Malaysian Economy in the Previous Two Decades? DOI: http://dx.doi.org/10.5772/intechopen.92214*

incorporate the leading macroeconomics indicators and expressed based on the ARDL framework:

$$\begin{aligned} \Delta lnY\_t &= a\_0 + \sum\_{i=0}^p q\_t \Delta lnY\_{t-i} + \sum\_{i=0}^p \theta\_t \,\Delta\_{\text{Speed-i}} + \sum\_{i=0}^p \beta\_t \,\Delta\_{\text{Leading-i}} + \delta\_1 \ln Y\_{t-1} \\ &+ \delta\_2 \,\text{Spread}\_{t-1} + \delta\_3 \,\text{Leading}\_{t-1} + \mathbf{v}\_t \end{aligned} \tag{2}$$

where *lnYt* is the economic growth indicated by industrial production index and expressed in natural logarithm, Spreadt is the yield spread between 10-year government bond and 3-month Treasury bill and Leadingt are the controlled variables for other macroeconomics leading indicators, *Δ* is first-difference operator, and p is the optimal lag length whereby the optimal lag length which represents the previous values, are being automatically selected based on Akaike info criterion (AIC). In consideration that the growth could be serially correlated, since previous growth might influence future growth, its past values are useful predictors themselves. This could also be the case for other independent variables, namely spread and leading.

The estimation model above will be applied onto three different samples, first on the whole sample (sample A) for the period of January 1996 to December 2016, while the second and third samples are based on the periods within the occurrences of the major crisis, from January 1996 to December 2000 (sample B) and from January 2007 to December 2009 (sample C), respectively. Our aim is to examine whether the long-run relationship among the variables, particularly the significance of the yield spread in explaining growth, still persists over different time periods.

The ARDL long-run form and bounds test is then undertaken for testing the existence of the long-run relationship, which is detected through the F-statistics (Wald test), and is said to be established if the F-statistics exceeds the critical value band, see Nkoro and Uko [32]. Specifically, the null hypothesis for no cointegration among variables in Eq. (2) is defined as H0: δ1 = δ2 = δ3 = 0 (where long-run relationship does not exist) against the alternative hypothesis of H1: δ1 6¼ δ2 6¼ δ3 6¼ 0 (long-run relationship does exist). Upon running the ARDL long-run form and bound test in Eviews 9.5, two sets of critical values are generated of which one set refers to I(0) and the other one refers to I(1). Critical values for the I(1) series are referred to as upper bound critical values, while the critical values for I(0) series are referred to lower bound critical values (Duasa, [31]). This is the bound testing procedure generated through the ARDL model and widely used in the estimation of long-run relationships when the properties of the time-series data are a mixture of I(0) and I(1).

If there is evidence of long-run relationship (cointegration), the following model is estimated:

$$
\ln Y\_t = \alpha\_1 + \sum\_{i=0}^p \eta\_t \ln Y\_{t-i} + \sum\_{i=0}^p \theta\_t \text{ Speed}\_{t-i} + \sum\_{i=0}^p \beta\_t \text{Leading}\_{t-i} + \mu\_t \tag{3}
$$

Subsequently, the ARDL specification of the short-run dynamics is derived by constructing the error correction model (ECM) of the following form:

$$
\Delta l n Y\_t = a\_2 + \sum\_{i=0}^p \varrho\_t \, \Delta l n Y\_{t-i} + \sum\_{i=0}^p \theta\_t \, \Delta\_{\text{Sprad}} + \sum\_{i=0}^p \beta\_t \, \Delta\_{\text{Leading}} + \mathfrak{w} \mathbf{e} \mathbf{m}\_t + \mathfrak{e}\_t \tag{4}
$$

where the ecmt is the error correction term and is defined as.

$$\text{Dec}m\_t = \ln Y\_t - a\_1 + \sum\_{i=0}^p \eta\_t \ln Y\_{t-i} + \sum\_{i=0}^p \theta\_t \text{ Speed}\_{t-i} + \sum\_{i=0}^p \beta\_t \text{Leading}\_{t-i} \tag{5}$$

period allows the testing on not only the significance of the yield spread on the growth, but on whether it is consistent throughout the periods, especially those during major crises. The optimal lag length is automatically selected by the ARDL method, based on the selection criteria of AIC as explained earlier. The calculated F-statistics for the cointegration test is displayed in **Table 2**. For sample A and B, the calculated F-statistics are above both the lower and upper bound critical values, leading to the rejection of the null hypothesis and to concur that there exists longrun relationship between the variables in the model. As for sample C, the calculated F-statistics is below the lower and upper bounds, which indicates that there is no cointegration, raising questions on the consistency of the relationship among the

*Has the Yield Curve Accurately Predicted the Malaysian Economy in the Previous Two Decades?*

The empirical results of the long-run relationship for sample A and B are presented in **Table 3**. It is interesting to note that during the longer time span of 20 years, the yield spread is not significant, as compared to a much shorter time

> **F-statistics value**

236 **4.4619** 2 10%

57 **5.2644** 2

36 0.8542 2

**Sample Dependent variable Independent Variables**

A *lnYt* 0.2179

B 0.09514\*

**Lag (k)**

**Significance level**

> 5% 1%

(0.3282)

(0.0207)

**Lower bound critical values I(0)**

> 2.63 3.10 4.13

**Spread Leading**

**Upper bound critical values I(1)**

> 3.35 3.87 5.00

1.0540 (0.9671)

0.0893\* (0.0349)

Next, the error correction model indicating the short-run dynamics is presented in **Table 4**, estimated for all of the samples A, B, and C. The lagged term of the yield spread appeared to be only significant during the crisis samples (B and C) but not for the whole 20-year period. A number of diagnostic tests for the error correction model are also applied where there is no evidence of serial correlation, heteroskedasticity, and ARCH (Autoregressive Conditional Heteroskedasticity) effect in the disturbances. All samples except for sample A passed the Jarque-Bera normality test, suggesting that the errors are normally distributed. Most importantly, the

span of 4 years and during the period of Asian financial crisis.

variables when the time period is altered.

*DOI: http://dx.doi.org/10.5772/intechopen.92214*

**Sample Data span Number of**

A January 1996 to December 2015

B January 1996 to December 2000

C January 2007 to December 2009

*F-statistics of cointegration relationship.*

**Table 2.**

*\**

**159**

**Table 3.** *Long-run model.*

*Significant at 1% level.*

**observations**

with all coefficients of the above short-run equation are equations relating to the short run dynamics of the model's convergence to equilibrium with *ψ* represents the speed of adjustments. The *ecmt* shows how much of the disequilibrium is being corrected, that is the extent to which any disequilibrium in the previous period is being adjusted in *Yt*.

## **5. Estimation results and discussions**

#### **5.1 Unit root tests**

Although ARDL cointegration technique does not require pre-testing for unit roots, this test is still carried out in order to avoid the ARDL model crash in the presence of integrated stochastic trend of I(2) (Nkoro and Uko [32]). In testing the nonstationarity of the series based on unit root test, there are two widely used tests by the econometricians, namely, the Augmented Dickey-Fuller (ADF) unit root test and Phillips-Perron (P-P) unit root test. Some important note on the presence or absence of unit roots, it helps to identify some features of the underlying data-generating process of a series. If a series has no unit roots, it is characterized as stationary, exhibiting mean reversion in that it fluctuates around a constant long-run mean. Alternatively, if the series feature a unit root, they are better characterized as nonstationary processes, which have no tendency to return to a long-run deterministic path, see Libanio [33].

The ADF test accounts for temporarily dependent and heterogeneously distributed errors by including lagged first differences of the dependent variable in the fitted regression. In contrast, the P-P test uses a nonparametric correction to take account for possible autocorrelation. The formula for the ADF test and the P-P test are not presented here, in consideration that it is a standard procedure in data inspection for stationarity. Nonetheless, a good explanation on the unit root stochastic process is covered by Nkoro and Uko [32]. In this study, the unit root test is testing based on ADF test. **Table 1** below presents the results from ADF indicating that there is a mixture of I(0) and I(1) with the absence of I(2) of the regressors, validating the use of ARDL in this analysis.

#### **5.2 Estimation results**

Subsequently, Eq. (2) is estimated on three samples of data, namely, sample A (January 1996 to December 2015), sample B (January 1996 to December 2000), and sample C (January 2007 to December 2009). This different estimation of the data


**Table 1.** *Unit root test.*

## *Has the Yield Curve Accurately Predicted the Malaysian Economy in the Previous Two Decades? DOI: http://dx.doi.org/10.5772/intechopen.92214*

period allows the testing on not only the significance of the yield spread on the growth, but on whether it is consistent throughout the periods, especially those during major crises. The optimal lag length is automatically selected by the ARDL method, based on the selection criteria of AIC as explained earlier. The calculated F-statistics for the cointegration test is displayed in **Table 2**. For sample A and B, the calculated F-statistics are above both the lower and upper bound critical values, leading to the rejection of the null hypothesis and to concur that there exists longrun relationship between the variables in the model. As for sample C, the calculated F-statistics is below the lower and upper bounds, which indicates that there is no cointegration, raising questions on the consistency of the relationship among the variables when the time period is altered.

The empirical results of the long-run relationship for sample A and B are presented in **Table 3**. It is interesting to note that during the longer time span of 20 years, the yield spread is not significant, as compared to a much shorter time span of 4 years and during the period of Asian financial crisis.

Next, the error correction model indicating the short-run dynamics is presented in **Table 4**, estimated for all of the samples A, B, and C. The lagged term of the yield spread appeared to be only significant during the crisis samples (B and C) but not for the whole 20-year period. A number of diagnostic tests for the error correction model are also applied where there is no evidence of serial correlation, heteroskedasticity, and ARCH (Autoregressive Conditional Heteroskedasticity) effect in the disturbances. All samples except for sample A passed the Jarque-Bera normality test, suggesting that the errors are normally distributed. Most importantly, the


#### **Table 2.**

*F-statistics of cointegration relationship.*


**Table 3.** *Long-run model.*


the relationship between slope of the yield curve (yield spread) and growth based on updated data and over a 20-year time period. Second, it is the first to employ the ARDL method on yield spread analysis, in consideration of the different order of

*Has the Yield Curve Accurately Predicted the Malaysian Economy in the Previous Two Decades?*

The empirical result proves the existence of a long-run relationship between the yield spread and growth in Malaysia. Though significant, the instability of the yield spread to affect the movement of growth in this analysis does not support the priori expectation on the predictive power of the yield curve, making it less reliable to be used as forecasting tool in the general economic condition. Further expansion of the local bond market in terms of issuance and trading may be the one of the keys to establish a much stronger relationship and predictive power of the yield curve in

This study is funded by the Research Acculturation Grant Scheme RAGS15-061-

Department of Economics, Kulliyyah of Economics and Management Science,

© 2020 The Author(s). Licensee IntechOpen. This chapter is distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/ by/3.0), which permits unrestricted use, distribution, and reproduction in any medium,

integration among the variables tested.

*DOI: http://dx.doi.org/10.5772/intechopen.92214*

**Acknowledgements**

**Author details**

**161**

Maya Puspa Rahman

International Islamic University, Malaysia

provided the original work is properly cited.

\*Address all correspondence to: mayapuspa@iium.edu.my

assessing the direction of the Malaysian economy.

0124, Ministry of Higher Education, Malaysia.

*1. t-statistic in parentheses; 2. Far is the F-statistic of Breusch-Godfrey serial correlation LM test. Farch is the F-statistic of ARCH Test. JBnormal is the Jarque-Bera Statistic of Normality Test.*

*\* Significant at 1% level.*

*\*\*Significant at 5% level.*

*\*\*\*Significant at 10% level.*

#### **Table 4.**

*Error correction model for all samples.*

significant of the error correction term (*ecmt*) for all samples also provide evidence of causality in at least one direction, with the negative coefficient indicating high rate of convergence to equilibrium. Nonetheless, the mixed signs of the yield spread do not match with the theory previously discussed, rendering difficulty in making general inferences with regard to the relationship of the yield spread to growth.

Though significant, the instability of the yield spread to affect the movement of growth in this analysis does not support the priori expectation on the predictive power of the yield curve, in Malaysia. This is consistent with the finding of Zulkhibri and Abdul Rani [10] that yield spread contains little information on the direction of the overall economy. As such, despite the fact that the bond market (both conventional and Islamic) has grown rapidly over the past two decades, the market still needs to deepen more in terms of issuance and trading, so as to facilitate the efficiency of yield curve movement, which could possibly be tagged along with the growth in the future.
